Shifting the focus from control to communication: The streams objects Environments model of communicating agents

6. Conclusions

Recent developments in Computing (e.g. networks, multimedia interfaces) have produces an extraordinary acceleration in the applicative needs for truly interactive, human-computer communication systems for various purposes. For instance, Education is an area of rapid growth of demand, Information another one. Object oriented technologies, considered to be an AI exclusive domain in the 70ties, are current state of the art and objects operate by exchanging messages. The shift from programs/algorithms within a single machine to processes/objects exchanging messages among various cooperating agents is perhaps the major current challenge in applicative Computing. It reduces basically to consider communication as driving control, not vice versa. The mental model of the underlying virtual machine is necessarily different from the previous one.

Human communication by exchanging messages shows that the pragmatic aspects of communication (goals, knowledge, intentions...) drive the success of the communicative processes. Emerging mainly from the needs of the AI community, in particular interoperability between knowledge bases for informative purposes, the Knowledge Sharing Effort has produced in the US a first mature language - called KQML - that incorporates pragmatic primitives, its performatives. From the needs of advanced educational applications, we have proposed a model and architecture called STROBE that shares the same objectives but suggests slightly different paths for achieving them.

Scheme is considered in STROBE both as a description and a prototyping language. Message exchange is viewed as a dynamic process where agents decide the next move by assembling / selecting it after evaluating the stream of previous messages exchanged with the partners. Agents are equipped with cognitive environments, i.e. trees of frames labeled according to the history and the partner's messages. Scheduling of tasks in agents occurs in a fashion similar to that of operating systems, but unlike them (and actors) is influenced by the explicit pragmatic layer that belongs to messages. When an agent processes a message, it is seen by its partners as an enhanced interpreter of KQML-like messages, that includes as a component an interpreter of the content language indicated in the message.

The integration of the proposed architecture with available networks, interfaces and platforms, i.e. most of the "lower level" technologies necessary for a realistic experimentation is achieved by coupling Scheme with Java.

Even if the work done is far from being complete, its properties become progressively clearer. One concrete result is that the concept is feasible. Another, to be debated, hypothesis is that it is also helpful for studying complex communicative phenomena by experimentation. Finally, the most important claim - yet only partially justified - is that it is relatively simple. If this were true, we would have reached our goal, as complexity of languages, tools and systems too often has hindered the accumulation of results in the research and a sensible utilization of them in practical applications.

7. References

[1] M. Woolridge, J. P. Muller, and M. Tambe, “Intelligent Agents II. Agent Theories, Architectures, and Languages,” in Lecture Notes in Artificial Intelligence, vol. 1037, J. G. Carbonell and J. Siekmann, Eds. Berlin: Springer Verlag, 1996, pp. XVIII- 437.

[2] J. Self, “Artificial Intelligence in Education: Towards Computational Mathetics,” University of Lancaster, Lancaster, UK, Unpublished draft May 1995 1995.

[3] S. A. Cerri and V. Loia, “A Concurrent, Distributed Architecture for Diagnostic Reasoning,” Journal of User Modeling and User Adapted Interaction, vol. 7, pp. 69-105, 1997.

[4] R. Kent Dybvig, The Scheme Programming Language; ANSI Scheme: Prentice-Hall, Inc., 1996.

[5] C. Queinnec, Les langages LISP. Paris: InterEditions, 1994.

[6] H. Abelson and G. J. Sussman, Structure and Interpretation of Computer Programs, 2nd ed. Cambridge, Mass.: MIT Press, 1996.

[7] J. Mayfield, Y. Labrou, and T. Finin, “Evaluation of KQML as an Agent Communication Language.,” in Intelligent Agents II. Agent Theories, Architectures, and Languages, vol. II, Lecture Notes in Artificial Intelligence, M. Woolridge, J. P. Muller, and M. Tambe, Eds. Berlin: Springer, 1996, pp. 347-360.

[8] S. A. Cerri, “Models and Systems for Collaborative Dialogues in Distance Learning,” in Collaborative Dialogue Technologies in Distance Learning, vol. 133, Computer and Systems Sciences, M. F. Verdejo and S. A. Cerri, Eds. Berlin: Springer Verlag, 1994, pp. 119-125.

[9] R. Pinkilton and C. Parker-Jones, “Interacting with Computer-Based Simulation: the role of dialogue,” Computers Educ., vol. 27, pp. 1-14, 1996.

[10] H. C. Bunt, “Dialogue analysis and speech act theory,” Institute for Perception Research, Eindhoven 330/II, 1978.

[11] S. Giroux, “Open Reflective Agents,” in Intelligent Agents II. Agent Theories, Architectures, and Languages, vol. II, Lecture Notes in Artificial Intelligence, M. Woolridge, J. P. Muller, and M. Tambe, Eds. Berlin: Springer, 1996, pp. 315-330.

[12] S. Leman, P. Marcenac, and S. Giroux, “A Generic Architecture for ITS based on a MultiAgent Model,” in Intelligent Tutoring Systems, vol. 1096, Lecture Notes in Computer Science, C. Frasson, G. Gauthier, and A. Lesgold, Eds. Montréal: Springer Verlag, 1996, pp. 75-83.

[13] J. Moeschler and A. Reboul, Dictionnaire encyclopédique de pragmatique. Paris: Editions du Seuil, 1994.

[14] J. R. Searle, Speech acts. An essay in the philosophy of language. Cambridge, UK: Cambridge University Press, 1969.

[15] D. Guin, “Towards Models of Interaction Between an Artificial Agent and Human one,” in Collaborative Dialogue Technologies in Distance Learning, vol. 133, ASI Series F: Computers and Systems Sciences, M. F. Verdejo and S. A. Cerri, Eds. Berlin: Springer Verlag, 1994, pp. 170-180.

[16] P. Reitz, “Contribution à l'étude des environments d'apprendissage. Conceptualization, spécification et prototypage.,” PhD Thesis. Departement d'Informatique. Montpellier: Université de Montpellier II, 1992.

[17] S. A. Cerri, “The "Natural Laboratory" Methodology Supporting Computer Mediated Generic Dialogues,” in Collaborative Dialogue Technologies in Distance Learning, vol. 133, Computer and Systems Sciences, M. F. Verdejo and S. A. Cerri, Eds. Berlin: Springer Verlag, 1994, pp. 181-201.

[18] J. Self, “Computational Viewpoints,” Computing Department, University of Lancaster, Lancaster, UK AI Report n. 44, March 1990.

[19] S. A. Cerri, “Cognitive environments in the STROBE model,” presented at EuroAIED: the European Conference in Artificial Intelligence and Education, Lisbon, Portugal, 1996.

[20] B. Cheikes, “Should ITS Designers Be Looking For A Few Good Agents?,” presented at AI-ED Workshop on Authoring Shells for Intelligent Tutoring Systems, Washington, DC, 1995.

[21] A. Paiva, “Learner Modelling Agents,” presented at EuroAIED: the European Conference in Artificial Intelligence and Education, Lisbon, Portugal, 1996.

[22] S. A. Cerri, “Computational Mathetics Tool kit: architecture's for dialogues,” in Intelligent Tutoring Systems, vol. 1096, Lecture Notes in Computer Science, C. Frasson, G. Gauthier, and A. Lesgold, Eds. Montréal: Springer Verlag, 1996, pp. 343-352.

[23] D. P. Friedman and A. Yonezawa, “Special issue on Computational Reflection.,” in LISP and symbolic computation. An International Journal., vol. 9, Nos. 2/3, R. R. Kessler, Ed. Dordrecht: Kluwer, 1996, pp. 151-257.

[24] S. Jefferson and D. P. Friedman, “A Simple Reflective Interpreter,” LISP and symbolic computation, vol. 9, pp. 181-202, 1996.

[25] K. Noermark, “Simulation of Object-Oriented Concepts and Mechanisms in Scheme,” Institute of Electronic Systems, Aalborg University, Aalborg, DK R 90 01, ISSN 0106-0791, 1991.

[26] S. C. Garrod and G. Doherty, “Conversation, co-ordination and convention: an empirical investigation of how groups establish linguistic conventions,” Cognition, vol. 53, pp. 181-215, 1994.

[27] S.-D. Lee and D. P. Friedman, “First-Class Extents,” Computer Science Departement, Indiana University, Bloomington, Indiana August 4, 1992.

[28] A. S. Fabiano and S. A. Cerri, “Concurrent, asynchronous search for the availability of knowledge,” Applied Artificial Intelligence Journal, vol. 10, pp. 145-161, 1996.

[29] J. Ferber, Les systèmes multi-agents. Vers une intelligence collective. Paris: InterEditions, 1995.

[30] Y. Labrou and T. Finin, “A semantics approach for KQML - a general purpose communication language for software agents.,” presented at Third International Conference on Information and Knowledge Management, 1994.

[31] R. Pinkilton, “Analyzing Educational Discourse: The DISCOUNT Scheme,” Computer-Based Learning Unit, The University of Leeds, Leeds, UK, TR n. 96, April 1996.

[32] P. Brusilowsky, E. Schwartz, and G. Weber, “ELM-ART: An Intelligent Tutoring System on World Wide Web,” in Intelligent Tutoring Systems, vol. 1096, Lecture Notes in Computer Science, C. Frasson, G. Gauthier, and A. Lesgold, Eds. Montréal: Springer Verlag, 1996, pp. 261-269.

[33] S. A. Cerri, “Learning Computing: understanding Objects by understanding Variables and Functions,” Object Currents: The first Online Hypertext Journal on Internet concerning Object Oriented Programming, vol. 1, pp. http://www.sigs.com/objectcurrents/, 1996.

[34] G. Agha, Actors: A Model of Concurrent Computation in Distributed Systems. Cambridge, Mass.: The MIT Press, 1986.

[35] S. A. Cerri, P. Mattijsen, and M. Van Dijk, “Computer Aided Design of Didactic Software,” in Interactive Techniques in Computer Aided Design, ACM-Italian-Chapter, Ed. Bologna, Italy: IEEE Computer Society, 1978, pp. 195-203.

[36] S. A. Cerri and J. Breuker, “A rather intelligent language teacher,” presented at Artificial Intelligence : Proc. AISB-80 Conference, Amsterdam, NL, 1980. 

[37] S. A. Cerri, A. Gisolfi and V. Loia, “Towards the Abstraction and Generalization of Actor-based Architectures in Diagnostic Reasoning ,” in: >

[38] G. Dionisi, “AL: un linguaggio per descrivere la comunicazione tra agenti ,” Tesi di laurea in Scienze dell’Informazione, Università di Milano, october 1998.

[39] D. Kafura and J.P. Briot, “Actors & Agents ,” IEEE Concurrency, 6,2, pp. 24-29,1998.

_{8. Acknowledgements}

I am particularly grateful to Antonio Gisolfi, who invited me several times to conferences organized by him in Ravello, with a great professionality, including the one where I presented an extension of this work. Vincenzo Loia is the only person that has succeeded in convincing me to dedicate efforts in writing scientific papers, in particular recently when many occasions in this Country presented clear evidences that performing experimental research in Computing may be often considered loosing your time. Erick Gallesio has contributed in a significant way to refining my knowledge of Scheme, and is co-author of our joint presentation at the VIM Conference in Ravello. Julian Padget, finally, was always supportive and patient, particularly in accepting my delays. Thank you.

 This paper is a revised version of the one appeared in the Proceedings of JFLA97, Journées francophones des langages applicatifs, Collection Didactique de l’INRIA, Marc Gengler et Christian Queinnec (eds.), pag. 145-168 (1997) with the title: A simple language for generic dialogues: “speech acts” for communication.

1 In Queinnec's book [5] , chapter 5 is entirely dedicated to the relations between Scheme, Denotational semantics and lambda calculus.

2 Model and architecture are used as synonyms: a computational model is the abstract view of a computational architecture, i.e. the set of expressions in a language, together with the underlying virtual machine (the evaluation method of the expressions) that describes the solution to a class of problems. We also share the view of those [6]  that believe that a program is a language for a class of problems. As an abstraction, STROBE is a model and an architecture. The (prototype) Scheme programs supporting STROBE functionality’s, together with the semantics of Scheme constitute a (prototypical) programming language embedded in Scheme.

3 Whether this effect is measurable and how, is an important, but still open issue.

4 Role exchange in Educational applications has been described in [15]  as an application of a model developed for machine learning reported in [16] . The NAT*LAB project reported also in [17]  was exploring experimentally the potential advantages of role exchanges in educational dialogues according to the Natural Laboratory methodology.

5 See: http://www.cs.umbc.edu/kqml/ for most KQML papers emerging from the Knowledge Sharing Effort.

6 This section is a revision from [19, 22]. As these Conferences were for the AI in Education community, it seems necessary for us to survey here at least a minimal information. Notice that STROBE was initially concerned with synchronous, two-human-partner communication, even if mixed initiative was allowed by introducing a third partner called coordinator.

7 Environment means here the set of frames that bind variables to values; i.e. the meaning is the one common in Programming Language research. In agent's research the meaning of environment may be associated to the set of stimuli external to the agent.

8 By introducing a third "coordinator" agent that decided, at the end of each exchange, which partner agent between the two will be allowed to take the initiative for the next exchange.

9 (called "environment" in the Agent's literature)

10 Concerning cognitive studies on shared meanings, we find in [26]  that cooperation between humans is established after a long process negotiating a common "vocabulary".

11 The notion of autonomous agent is controversial (cf. [29]  for a comprehensive presentation). The minimal level of autonomy that dialogues require is the one that allows any agent to take spontaneously the initiative. That may occur asynchronously. An agent, even processing another agent's query, may feel the need to put a query on turn, in order to use the results of the last one for the benefit of assembling an answer to the first one. This observation allows deducing that scheduling of activities within agents should occur internally to the agent. That would be excluded if agents would communicate with each other according to queued messages, where scheduling of message processing is external to the single agent and proprietary of the shared communication language. Actors, e.g., do have this common scheduling property and therefore we suspect that they may not easily be applied to modeling autonomous agent's dialogues.

12 Even if KQML papers refer to "cognitive states of agents" [30]  it is clear that the "cognitive" property of KQML agents refers only to software agents, e.g. knowledge bases, assumed implicitly to be consistent and persistent.

13 STk 3.0 is freely available via anonymous ftp at ftp.unice.fr . It has been developed and is currently maintained by Eric Gallesio: eg@ unice.fr . The current version available is 3.99 .

14 These are: active-socket for creating an active socket; wait-input-socket for waiting until at least a socket receives a message, passive-socket to create a passive socket awaiting requests for connection, passive-socket-accept that accepts a connection request on a passive socket and restitutes a new socket to be used for communication; select-input-socket selecting which sockets do have available messages. The current version of STk supports sockets nicely.

15 At the time of writing, a third implementation of the STROBE architecture is available with the name AL (Alice Language) [38]. In AL there is a reduction of primitive performative types to three, while any other one may be constructed from these by designing the behavior associated to new performatives in the form of a finite state automaton associated to each agent. Experiments with AL are currently ongoing, mainly in the area of electronic commerce. AL is freely available on request.

16 As this paper is a revised version of one written earlier, most of the enhancements have been realized and are going to be published in due time.

17 In [9] , for instance, two experiments were performed: the first analyzed protocols from experimenter - learner interactions, including the common phenomenon of mixed initiatives, and the second from experimenter - student1 - student2 interactions. From their results, even a "simple" statistical analysis of the types of moves used in the two different dialogues may allow to draw conclusions concerning the learning outcomes, that may be used as requirements for subsequent applications.

18 The solution is the one commonly used when accessing shared variables in Distributed DBMS.

19 Remarkably, partner-labeled environments are proposed also in important applicative situations where WWW servers manage interactions with clients interleaved with each other and requiring non purely functional computations on the server (e.g. using a CL-HTTP server or MIT Scheme server remotely, as shown in [32]  and also in [33] ). That was in fact the architecture of the PLATO system (Un. Illinois, then CDC) in the 60ties, when PLATO was able to manage in real time up to 500 - 1000 simultaneous users (students and teachers!) with a centralized computing power of the order of a few MIPS. The historical note is not folklore, because current developments in the WWW will require substantial progress in "centralized" operating systems supporting threads of communications with users, if WEB computers will become diffused. A trend that is opposite to the one, typical of the 80ties, that consisted in distributing computing power and control to PCs and workstations.

20 At the moment of writing, the second prototypical language implementing the STROBE architecture is available with the name JASKEMAL. The scheduler of messages in JASKEMAL is indeed proprietary of the agent, thus offers the opportunity to build really autonomous agents, each behaving as an operating system with a dynamic, proprietary scheduler.

21 Didactic Augmented Recursive Transition Networks.

22 While a description of a dialogue among autonomous agents may hardly be described by a single ATN or a single Petri Net (see, for example, [29] ) – a result already available in [35, 36] – we believe that each agent may suitably be described by a finite state automaton FSA (or an ATN or a Petri Net). We agree with [38] that a FSA is simpler as a Petri net, even if an ATN has more representational power, as it is an extension of a FSA with sub-nets and actions on registers.